论文解读
音频幻觉为何总说“有”：用静默对照与首步门控校正肯定偏差的 TAD
针对二元音频问答中模型因语言先验而虚报存在的幻觉，TAD 以静默音频为对照做对比解码并仅在首步用音频增益门控惩罚肯定词，在 AudioCaps-Hallucination 与 Clotho-AQA 上提升拒绝能力但在部分设置下以精度与准确率下降为代价。
音频问答
论文解读
猜对不算懂：用选项重排伪集成让音乐问答模型学会拒答
针对音乐音频语言模型在四选一问答中被迫作答而无法表达不知道的问题，论文用同一 TinyMU 检查点经答案不变扰动构造伪集成，主证据是四种选项排序平均将准确率从 55.7% 提升到 59.2% 并将误差保留曲线面积从 0.293 降到 0.261，代价是每次提问需要 4 次前向传播且无需重新训练。
论文解读
从听见写入选项到用选项做决定：音频大模型中声学证据的两阶段路径
在必须依赖声音的选择题上训练能提升准确率并同步增强对真实音频的行为敏感度，其内部机制是早期到中层先让音频塑造选项表征、再在中层到晚层强化选项对最终答案的贡献，而可学习的 LoRA 更新在模型特定的层带上起决定作用。
论文解读
ALM2Vec: Learning Audio Embeddings for Universal Audio Retrieval with Large Audio-Language Models
音频检索 | 7.4/10
论文解读
AOR-Bench: Do Large Audio Language Models Over-Refuse Pseudo-Harmful Queries?
音频问答 | 7.9/10
论文解读
Learning from Audio-Dependency Errors: Data Curation Strategies Based on Model Confusion Patterns in Audio Question Answering
音频问答 | 6.9/10
论文解读
EChO-Agent: Evidence Chain Orchestration Agent for Audio Reasoning
音频问答 | 6.1/10
论文解读
AudioDER: A Deduplication-Enhanced Reasoning Dataset for Post-Training Large Audio-Language Models
音频问答 | 7.3/10
论文解读
Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources
音频问答 | 6.9/10
论文解读
Beyond Text Following: Repairable Arbitration Reversals in Audio-Language Models
音频问答 | 6.4/10
论文解读
MusTBENCH: Benchmarking and Advancing Temporal Grounding in Music LLMs
音乐生成 | 7.5/10
论文解读
Audio-Mind: An Auditable Agentic Framework for Audio Understanding
音频问答 | 8.7/10
论文解读
VoiceGiraffe: A Benchmark for Extreme Long-Context Audio-Language Understanding
音频问答 | 7.0/10
论文解读
A Survey of Audio Reasoning in Multimodal Foundation Models
音频推理 | 7.7/10
论文解读
Heterogeneity-Aware Dataset Scheduling for Efficient Audio Large Language Model Training
音频问答 | 7.0/10
论文解读
PlanRAG-Audio: Planning and Retrieval Augmented Generation for Long-form Audio Understanding
长音频理解 | 7.4/10
论文解读
Heterogeneity-Aware Dataset Scheduling for Efficient Audio Large Language Model Training
音频问答 | 7/10
论文解读
Can Large Audio Language Models Ignore Multilingual Distractors? An Evaluation of Their Selective Auditory Attention Capabilities
音频问答 | 6.5/10
论文解读
Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs
模型评估 | 8.0/10
论文解读
Omni-DeepSearch: A Benchmark for Audio-Driven Omni-Modal Deep Search
基准测试 | 6.0/10